AI agent accessed Anthropic internal data

Cathedral, a cybersecurity company founded by former Department of Government Efficiency (DOGE) members, says researchers used AI agents to find vulnerabilities in third-party software used by Anthropic in under 12 hours and gained access to some of its internal systems. According to its preliminary report, potentially accessible information includes meeting records, compensation data and Slack discussions involving Anthropic's advanced models. Anthropic said the researchers first spent several days attacking directly without success, then turned to a vulnerable third-party business application, and that the under-12-hours figure refers only to the successful phase; what was touched was a limited scope of business, human resources and IT systems, but users' conversations with Claude, model responses, source code and trained model parameters were not viewed, and the company cut off access the same day and disabled the relevant software. The two sides interpret it differently: Cathedral calls it a warning about the growing attack capabilities of AI, while Anthropic says it exaggerates national security implications and that not all activities were reported under the rules of its vulnerability disclosure program. Public information so far does not indicate that Anthropic's models were compromised.